The jump in global temperatures in September 2023 is extremely unlikely due to internal climate variability alone (2024)

The observed record margin is a rare event in the climate model simulations

We argue that internal climate variability alone is unlikely to explain the unusually large margin by which September’s record was broken. To illustrate this, we consider simulations from three climate model ensembles: Coupled Model Intercomparison Project 6² (CMIP6), the 100-member Max Planck Institute Grand Ensemble³ (MPI-GE), and the 100 member Community Earth System Model version 2⁴ (CESM2-LE). These are well-established models known for their reliable simulations of both internal climate variability, such as the El Niño-Southern Oscillation, and the forced response to greenhouse gas forcing.

By looking at each model simulation for the period 1970–2050 and searching for the margins by which the monthly records are broken in each simulation, we obtain a total of 5166 September records in CMIP6 models, 1431 in MPI-GE and 2068 in CESM2-LE (see Methods). These distributions are shown in Fig. 2a–c. The distribution of record margins results from the unforced internal variability and the forced greenhouse gas-induced trend. A larger trend or higher variability, or both, increases the likelihood of large record margins.

We briefly analyse the magnitude of internal variability in the models and observations by calculating the standard deviation of the detrended September mean temperature. Especially in CMIP6, the model-simulated standard deviation of September temperatures tends to be larger than in the observations (Supplementary Fig. 1). This suggests that the internal variability of the models is at least not smaller than in the observations, and thus the probabilities are not being underestimated.

The observed margin from September 2023 (0.5 °C) is shown as a black dashed line in Fig. 2. As can be seen, the observation falls in the far right tails of the model simulated margins. In CMIP6 (Fig. 2a), only three of the 5166 model-simulated margins exceed the observed margin, corresponding to the 99.94th percentile of the model distribution. In MPI-GE (Fig. 2b) the observation is completely outside the distribution, and in CESM2-LE (Fig. 2c) there is only one margin higher than the observation, meaning that the corresponding percentile is 99.95%.

We calculate the probability of the observed margin from the fitted Generalized Extreme Value (GEV) distribution. In the fit, we consider only the most extreme margin from each simulation, similar to the observations (see Methods). For CMIP6, we obtain a p-value of 0.004 (Fig. 2d, see Supplementary Table 1 for the confidence intervals). For MPI-GE and CESM2-LE, the p-values are 0.018 and 0.01 (Fig. 2e, f), respectively. These values are generally consistent with the empirically sampled probabilities, the probabilities for the 1990–2050 period and the probabilities for the August–October period (Supplementary Fig. 2, Supplementary Table 1).

We repeat the analysis for September by excluding those climate models that, based on the Hausfather et al.⁵ analysis, fall outside the likely transient climate response (66% probability range) of 1.4–2.2 °C. This further reduces the p-value for the CMIP6 models, giving a result of p = 0.002 (Supplementary Table 1).

Furthermore, the discrepancy between the observed and simulated margins is almost equally striking when all calendar months are considered by the models. Considering all months, the p-values of the observed margin are 0.029 in CMIP6, 0.017 in MPI-GE and 0.025 in CESM2-LE (Supplementary Table 1). This is despite the fact that the internal variability of the climate is greater in the northern hemisphere (NH) winter months than in the NH summer months, as El Niño tends to peak in the NH winter. Therefore, the margins for breaking records are generally greater in the NH winter months than in the NH summer months.

For comparison, we also briefly examined the probability of the observed record margin of 0.47 °C in February 2016. In this case, we obtained p-values of 0.115 for CMIP6, 0.078 for MPI-GE and 0.141 for CESM2-LE. The margin observed in February was therefore about an order of magnitude more likely than the one observed in September, and thus more likely to be due to internal variability alone. However, it is worth noting that in February 2016, super El Niño had just peaked, when its impact on global temperature was near maximum. This is not the case for September 2023.

The most plausible explanation for the model-observation discrepancy in September 2023 would be that the observed combination of forced warming and internal variability is so rare that it does not occur in the models. The strengthening El Niño following a triple La Niña event observed in 2020–2022 has occurred only a few times since 1950, and not earlier in the 21st century⁶. However, large ensemble models are designed to capture such rare climate anomalies, and we found that no member in MPI-GE and only one member in CESM2-LE simulated temperature jumps as large as the one observed in September 2023.

It is also worth noting that increased solar activity may have contributed to the record margin in September 2023. However, solar forcing is included in CMIP6 models⁷, so while it may have added a few hundredths of a degree to the record margin, it is unlikely that increased solar activity contributed to the model-observation discrepancy, although the solar cycle 25 may have risen slightly faster than the estimate prescribed in the scenario.